It was a very busy month, this April in the AI
| Date | Lab / Organisation | Type | Release / Event | Key Fact |
|---|---|---|---|---|
| 2 apr | Model | Gemma 4 | 4 sizes (E2B, E4B, 26B MoE, 31B Dense), Apache 2.0, 256K context, 31B beats models 20x its size | |
| 2 apr | Alibaba | Model | Qwen 3.6-Plus | Proprietary flagship, 1M context, native multimodal, agentic coding |
| 6 apr | Anthropic | Deal | Google + Broadcom uitbreiding | Multi-gigawatt TPU-uitbreiding; deel van $50B US compute commitment uit nov 2025 |
| 7 apr | Z.ai (Zhipu) | Model | GLM-5.1 | 754B MoE / 40B active, MIT, #1 SWE-Bench Pro at 58.4%, getraind op Huawei Ascend chips |
| 7 apr | Anthropic | Model | Claude Mythos Preview | Beperkt tot ~50 partners via Project Glasswing, $100M credit pool, vond 27-jaar oude OpenBSD bug |
| 8 apr | Meta | Model | Muse Spark | Eerste model uit Meta Superintelligence Labs (Alexandr Wang); proprietary; vervangt Llama-merk |
| 13 apr | MiniMax | Model | M2.7 | Self-evolving capabilities — model update zijn eigen weights via productie feedback loops |
| 13 apr | Moonshot AI | Model | Kimi Code K2.6 (preview) | Coding-gerichte preview vooruitlopend op K2.6 full release |
| 14 apr | Anthropic | Bestuur | Vas Narasimhan benoemd in LTBT | Long-Term Benefit Trust breidt uit; Mike Krieger stapt op uit Figma board |
| 16 apr | Anthropic | Model | Claude Opus 4.7 | Verslaat Opus 4.6, GPT-5.4 en Gemini 3.1 Pro op key benchmarks; 1M context, betere vision (3.75 MP), nieuwe tokenizer, “task budgets” feature |
| 16 apr | Alibaba | Model | Qwen 3.6-35B-A3B | Open source, Apache 2.0, 35B params / 3B active |
| 17 apr | xAI | Model | Grok 4.3 Beta | Native video input, 2M context (grootste in Westerse closed models), $300/mo SuperGrok Heavy |
| 17 apr | Anthropic | Product | Claude Design (Anthropic Labs) | Visuele prototyping tool aangedreven door Opus 4.7; concurrent voor Figma/Canva; research preview |
| 20 apr | Amazon | Deal | $25B Anthropic investering | $5B direct + $20B contingent op $380B waardering. Anthropic commits $100B+ aan AWS over 10 jr, tot 5 GW Trainium2/3/4 capaciteit |
| 20 apr | Moonshot AI | Model | Kimi K2.6 | 1T MoE, Modified MIT, 262K context, 300 sub-agents in parallel (“Swarm 2.0”), 13u autonoom coderen, 58.6% SWE-Bench Pro |
| 20 apr | Alibaba | Model | Qwen 3.6 Max Preview | Proprietary, top scores op 6 coding benchmarks (SWE-bench Pro, Terminal-Bench 2.0, etc.) |
| 22 apr | Alibaba | Model | Qwen 3.6-27B | Open source mid-size coding model, op Hugging Face / ModelScope |
| 23 apr | xAI | Model | grok-voice-think-fast-1.0 | #1 op τ-voice Bench at 67.3%, verslaat Gemini en GPT Realtime, 25+ talen, draait Starlink phone sales |
| 23 apr | OpenAI | Model | GPT-5.5 (“Spud”) | Eerste volledige retrain sinds GPT-4.5, 1M context, $5/$30 per M tokens, 6 weken na 5.4; ook GPT-5.5 Pro |
| 23 apr | Anthropic + NEC | Deal | Strategische Japan-partnerschap | NEC eerste Japan-based global partner; Claude voor ~30.000 NEC medewerkers; sector-specifieke AI voor finance, manufacturing, lokale overheid |
| 24 apr | DeepSeek | Model | DeepSeek V4 (preview) | V4-Pro: 1.6T / 49B active (grootste open-weight model); V4-Flash: 284B / 13B active; MIT, 1M context, $1.74/$3.48 per M (Pro), Huawei Ascend 950 + Cambricon |
| 24 apr | Deal | $40B Anthropic investering | $10B direct op $350B waardering + $30B contingent. Plus 5 GW TPU-capaciteit over 5 jaar | |
| 24 apr | Anthropic | Beleid | Election safeguards update | Beleidsupdate over Claude’s rol bij US midterms en wereldwijde verkiezingen 2026 |
| 27 apr | Anthropic | Uitbreiding | Sydney kantoor opent | Theo Hourmouzis benoemd als GM Australië & Nieuw-Zeeland |
*Bronnen: Anthropic, OpenAI, Meta, Google, Bloomberg, CNBC, TechCrunch, Reuters, Al Jazeera, Wikipedia, MIT Sloan Review, Axios, GeekWire, Fortune, VentureBeat.